From Product Hunt
MCP tools and AI memory solutions dominate today. Multiple launches around giving AI assistants better context and control.
MemoryPlugin — Persistent Memory Across AI Tools
Browser extension that solves AI amnesia. Works across 17+ platforms (ChatGPT, Claude, Gemini, Grok). Smart filtering saves 90% of tokens, Memory Buckets keep contexts separate. MCP integration available.
Yavy — Turn Any Website Into an MCP Server
Crawls websites, creates semantic search indexes, serves content via MCP protocol. AI assistants get real answers from real content. Supports live sync, zero-config setup, OAuth 2.1 for private sources.
Relay.app Agents — Build AI Agent Teams
AI automation platform for building teams of AI agents that work together. Allows non-technical users to create automated workflows. Focus on "building an AI team that works for you."
Ray 3.0 — Output Window for AI Agents
Debugging tool by Spatie now with MCP server support. AI agents can read/write directly to Ray. 60% less memory usage, archive feature for comparing debug sessions. Works with Laravel, JavaScript.
screenpipe — Open Source Local AI Screen Recorder
Continuously captures screen and audio, creating searchable memory of everything on your computer. OCR text extraction, audio transcription, natural language search. Data stays local. MIT licensed, cross-platform.
Moltcraft — Pixel World Dashboard for AI Agents
Open-source visual dashboard showing AI agents as pixel characters. Replace log-scrolling with visual monitoring. Click to chat, voice interaction. Pure HTML/CSS/JS, ~2MB, MIT licensed.
From Reddit
Big benchmark news with GPT 5.2 crushing ARC-AGI. Plus Alibaba drops a new coding model, METR evals show impressive agent horizons, and a prompt injection warning for anyone building agents.
New SOTA on ARC-AGI-1 at 90.5% ($11.64/task) and 54.2% on ARC-AGI-2. A year ago o3 scored 88% at $4.5k/task — this represents ~390X efficiency improvement.
Using GPT-5.2 X-High, scores as high as 75% at under $8/problem observed on ARC-AGI-2. The benchmark is approaching saturation.
Alibaba Releases Qwen3-Coder-Next
80B parameter model with only 3B active (MoE). Non-thinking mode focused. Tech report covers tool call generalization and reward hacking findings. Running well on AMD Strix Halo via ROCm. NVFP4 quantization at 45GB.
METR: Gemini 3 Pro Has 4-Hour Time Horizon
METR evaluations show Gemini 3 Pro can complete 50% of tasks that take humans 4 hours. At 80% success rate, slightly beats Claude Opus 4.5. Gemini 3.0 Pro still not GA; post-training improvements coming.
Beta Tester Hints at Claude Image Generation
Code strings referencing image creation/editing spotted in Claude builds. Anthropic may be preparing image generation capabilities. Could release alongside Claude Sonnet 5 and interactive widgets.
ACE-Step 1.5 — Open Source Music Generation
Performance close to Suno, MIT licensed. Full song in ~2 seconds on A100, under 10s on RTX 3090 (~4GB VRAM). Supports LoRAs, cover/repainting features. ComfyUI integration available. Best open-source music model yet.
Quick hits — security warnings and regulatory news.
Security Warning: Wallet-Drain Prompt Injection on Moltbook
Malicious post with "SYSTEM OVERRIDE" payload attempting to hijack agent wallet permissions. Warning for anyone building agents that browse social feeds with tool access.
X Offices Raided in France, UK Opens Grok Investigation
French authorities raided X offices while UK opens fresh investigation into Grok AI. Regulatory pressure mounting on Musk's AI and social media operations.
From Reddit
While everyone waits for Z-Image Edit weights, Meituan drops an alternative. Plus ComfyUI gets motion capture and a new vision-language model.
LongCat — Z-Image Edit Alternative Available Now
Meituan releases LongCat-Image-Edit while everyone waits for Z-Image Edit. Supports global/local editing, text modification, and reference-guided editing. Now has 8-step Turbo version (~29GB).
ComfyUI-Yedp-Mocap — Real-Time Motion Capture
Moves pose/hand/face detection from GPU to browser via MediaPipe. Real-time feedback before recording, saves VRAM for generation. Includes 3D pose viewer.
ComfyUI-Youtu-VL — Tencent's Vision-Language Model
4B Vision-Language Model with VLUAS training comes to ComfyUI. Lightweight but powerful for image understanding tasks.
From Reddit
Multiple window managers for Mac (pick your poison), a long-awaited per-app volume control, and self-hosted media analytics.
Native macOS Window Manager with Magnetic Snap
New window manager that doesn't require memorizing 50 keyboard shortcuts. Features "Magnetic Snap" mode — snap between windows with 1/3, 1/2, 2/3 width options. Currently in beta.
Tracearr — Analytics for Plex, Jellyfin, Emby
Real-time monitoring with account sharing detection. Mobile apps launching soon. Push notifications, kill streams from phone. Free, self-hosted, AGPL-3.0 licensed.
How Vibe Coding Is Killing Open Source
Hackaday article examining how AI-assisted "vibe coding" impacts open source development. Thoughtful piece on the changing landscape of OSS contributions in the AI era.
Quick hits — utilities and side projects.
AppVolume — Per-App Volume Control for macOS
Control each app 0-200%, mute with one click. The macOS feature that Windows has had forever. Early access, free to try.
MacsyZones 2.1 — Free Open Source Window Manager
v2.1 brings new features and better UX. The "cutest" Mac window manager, donation-supported.
Liqoria — Mac Music Player with Animated Wallpapers
Liquid Glass design with animated wallpapers synced to playback. Supports Apple Music, Spotify, YouTube. Native macOS with widgets.
Kahani — Audiobook Player for Audiobookshelf
Offline-first Android player for self-hosted audiobooks. Clean interface, forked from existing OSS with major facelift.